def get_seed_rule(tried_seed_rules, kernel_set, gl):
import utils, asp, functs, excps
seed_model, seed_rule, found_seed = None, None, False
str_tried_seed_rules = [r.as_string for r in tried_seed_rules]
for c in kernel_set:
utils.analyze_use_try([c], [], incremental_search=True)
(_, pos, negs,
optimization_score) = asp.ind(kernel_generalization=True,
incremental_search=True)
(all_solutions, _, _) = asp.ind(find_all_optimal=True,
incremental_search=True,
opt=optimization_score)
for solution in all_solutions:
if len(solution) > 1:
(use_2, _) = functs.split_use_2_3(solution)
(ok,
use_head_body_map) = functs.head_body_use_atoms_filter(use_2)
if ok:
new = utils.form_new_clauses(
use_head_body_map, incremental_kernel_search=True)
if not new[0].as_string in str_tried_seed_rules:
seed_model = use_2
seed_rule = new[0]
found_seed = True
break
else:
msg = 'Found a solution use(i,j) atom with no corresponding use(i,0) atom'
raise excps.Use_2_HeadNotAbducedException(msg, gl.logger)
if found_seed:
break
return (seed_model, seed_rule, pos, negs, optimization_score)
def generate_all_seed_rules(kernel_set, gl):
import utils, asp, functs, excps
generated_seeds, str_generated_seeds = [], []
for c in kernel_set:
for seed in generated_seeds:
utils.analyze_use_try([c], [],
incremental_search=True,
preserve=generated_seeds)
(_, pos, negs,
optimization_score) = asp.ind(kernel_generalization=True,
incremental_search=True)
(all_solutions, _, _) = asp.ind(find_all_optimal=True,
incremental_search=True,
opt=optimization_score)
for solution in all_solutions:
if len(solution) > 1:
(use_2, _) = functs.split_use_2_3(solution)
(ok, use_head_body_map
) = functs.head_body_use_atoms_filter(use_2)
if ok:
new = utils.form_new_clauses(
use_head_body_map, incremental_kernel_search=True)
if not new[0].as_string in str_generated_seeds:
generated_seeds.append(new[0])
str_generated_seeds.append(new[0].as_string)
else:
msg = 'Found a solution use(i,j) atom with no corresponding use(i,0) atom'
raise excps.Use_2_HeadNotAbducedException(
msg, gl.logger)
print('seeds:')
utils.see(generated_seeds)
#return (seed_model,seed_rule,pos,negs,optimization_score)
return generated_seeds
def generate_all_seed_rules(kernel_set,gl):
import utils,asp,functs,excps
generated_seeds,str_generated_seeds = [],[]
for c in kernel_set:
for seed in generated_seeds:
utils.analyze_use_try([c],[],incremental_search=True,preserve=generated_seeds)
(_,pos,negs,optimization_score) = asp.ind(kernel_generalization=True,
incremental_search=True)
(all_solutions,_,_) = asp.ind(find_all_optimal=True,
incremental_search=True,
opt=optimization_score)
for solution in all_solutions:
if len(solution) > 1:
(use_2,_) = functs.split_use_2_3(solution)
(ok,use_head_body_map) = functs.head_body_use_atoms_filter(use_2)
if ok :
new = utils.form_new_clauses(use_head_body_map,incremental_kernel_search=True)
if not new[0].as_string in str_generated_seeds:
generated_seeds.append(new[0])
str_generated_seeds.append(new[0].as_string)
else:
msg = 'Found a solution use(i,j) atom with no corresponding use(i,0) atom'
raise excps.Use_2_HeadNotAbducedException(msg,gl.logger)
print('seeds:')
utils.see(generated_seeds)
#return (seed_model,seed_rule,pos,negs,optimization_score)
return generated_seeds
def get_seed_rule(tried_seed_rules,kernel_set,gl):
import utils,asp,functs,excps
seed_model,seed_rule,found_seed = None,None,False
str_tried_seed_rules = [r.as_string for r in tried_seed_rules]
for c in kernel_set:
utils.analyze_use_try([c],[],incremental_search=True)
(_,pos,negs,optimization_score) = asp.ind(kernel_generalization=True,
incremental_search=True)
(all_solutions,_,_) = asp.ind(find_all_optimal=True,
incremental_search=True,
opt=optimization_score)
for solution in all_solutions:
if len(solution) > 1:
(use_2,_) = functs.split_use_2_3(solution)
(ok,use_head_body_map) = functs.head_body_use_atoms_filter(use_2)
if ok :
new = utils.form_new_clauses(use_head_body_map,incremental_kernel_search=True)
if not new[0].as_string in str_tried_seed_rules:
seed_model = use_2
seed_rule = new[0]
found_seed = True
break
else:
msg = 'Found a solution use(i,j) atom with no corresponding use(i,0) atom'
raise excps.Use_2_HeadNotAbducedException(msg,gl.logger)
if found_seed:
break
return (seed_model,seed_rule,pos,negs,optimization_score)
def incremental_solve(input_program,prior_theory_,**kwargs):
analyze_use_try(input_program,prior_theory_)
can_do_better = True
# get an initial model:
if not 'opt' in kwargs:
(initial_model,optimization) = asp.ind(kernel_generalization=True,
incremental_solve=True,
init_model=True)
else:
(initial_model,optimization) = asp.ind(kernel_generalization=True,
incremental_solve=True,
improve_model=True,opt=kwargs['opt'])
if initial_model != 'UNSATISFIABLE':
(solution,use_2,use_3) = form_program(initial_model,incremental_solve=True)
while can_do_better:
print('optimization: %s'%(str(optimization)))
optimization -= 1
if optimization == 28:
stop = 'stop'
analyze_use_try(solution,prior_theory_)
(better_model,_) = asp.ind(kernel_generalization=True,
incremental_solve=True,
improve_model=True,opt=optimization)
if better_model == 'UNSATISFIABLE':
# There are two cases here: Either we reached the true optimal
# theory size, or we took a path (with the initial "seed" model) that
# cannot be further improved, while in fact better hypotheses
# are possible. So first try to back-jump to the initial Kernel Set,
# to see if we can find a better "seed" model of the current size.
# If that fails, then we arrived at the optimum.
return incremental_solve(input_program,
prior_theory_,
opt=optimization,
last_model=solution,
use2model=use_2,
use3model=use_3)
can_do_better = False
else:
(solution,use_2,use_3) = form_program(better_model,incremental_solve=True)
else:
solution = kwargs['last_model']
use_2 = kwargs['use2model']
use_3 = kwargs['use3model']
"""
===================================================
Remember to update supports for the final solution
===================================================
"""
print('Found optimal')
return (solution,use_2,use_3)
def search_kernel_by_subsets(prior_theory_):
var_kernel = gl.current_var_kernel
found_solution = False
while not found_solution:
for i in range(6,len(var_kernel)+1):
i_subsets = itertools.combinations(var_kernel, i)
subset_counter = 0
for subset_ in i_subsets:
subset = list(subset_)
print('Searching #%s kernel %s-subset:'%(str(subset_counter),str(i)))
analyze_use_try(subset,prior_theory_)
#analyze_use_try(gl.current_var_kernel,prior_theory_)
out = asp.ind(kernel_generalization=True,incr_kernel_search=True)
print(out)
if out != 'UNSATISFIABLE':
(use_2,use_3) = functs.split_use_2_3(out)
(ok,use_head_body_map) = functs.head_body_use_atoms_filter(use_2)
if ok :
new = form_new_clauses(use_head_body_map)
else:
msg = 'Found a solution use(i,j) atom with no corresponding use(i,0) atom'
raise excps.Use_2_HeadNotAbducedException(msg,gl.logger)
found_solution = True
subset_counter += 1
if found_solution:
break
if not found_solution:
print('Failed to find a solution')
sys.exit()
def incremental_search(**kwargs):
import utils, core, asp, functs, excps
gl = core.global_vals
gl.use_dict = {}
is_new_example = kwargs['is_new_example']
retained, new, refined = kwargs['retcl'], kwargs['newcl'], kwargs['refcl']
prior_theory = [x for y in [retained, new, refined] for x in y]
target_score = gl.current_example_object.positive_count
if is_new_example:
utils.generate_kernel(**kwargs)
# revise the clauses in the prior theory one by one first
for c in prior_theory: # TODO
pass
var_kernel = gl.current_var_kernel
# =====================================
generate_all_seed_rules(var_kernel, gl)
# =====================================
found_solution = False
#current_opt_score = float('Inf')
#pos_covered = 0
#negs_covered = float('Inf')
best_score = 0
tried_seed_rules = []
while not found_solution:
partial_rules, tried_kernel_clauses = [], []
kernel_loop_counter = 0
for c in var_kernel:
if 'terminatedAt' in c.as_string:
stop = 'stop'
tried_kernel_clauses.append(c)
partial_rules.append(c)
utils.analyze_use_try(partial_rules, [],
incremental_search=True)
if kernel_loop_counter == 0:
(model, seed_rule, pos, negs,
optimization_score) = get_seed_rule(
tried_seed_rules, var_kernel, gl)
tried_seed_rules.append(seed_rule)
print('New seed rules: ')
for r in tried_seed_rules:
print(r.as_string)
else:
(model, pos, negs,
optimization_score) = asp.ind(kernel_generalization=True,
incremental_search=True)
partial_rules.extend(tried_seed_rules)
#(all_solutions,pos,negs) = asp.ind(find_all_optimal=True,
# incremental_search=True,
# opt=optimization_score)
#new_score = pos - negs - average_size(all_solutions)
new_score = pos - negs - average_size_(model)
if new_score > best_score:
best_score = new_score
print('pos,negs,average_size,best_score: ', pos, negs,
average_size_(model), best_score)
#if pos >= pos_covered and negs <= negs_covered:
if False:
pos_covered, negs_covered = pos, negs
else:
# At this point, including any one solution from all_solutions reduces
# the current score. Hence, we will use one such solution as a new
# addition and search each tried kernel clause again, in order to retrieve
# the previous score (or even improve it)
#new_addition,added = [],[]
"""
for solution in all_solutions:
if len(solution) > 1:
(use_2,_) = functs.split_use_2_3(solution)
(ok,use_head_body_map) = functs.head_body_use_atoms_filter(use_2)
if ok :
new = utils.form_new_clauses(use_head_body_map,incremental_kernel_search=True)
for n in new:
if not n.as_string in added:
new_addition.append(n)
added.append(n.as_string)
"""
(use_2, _) = functs.split_use_2_3(model)
new = []
if len(use_2) > 1:
(ok, use_head_body_map
) = functs.head_body_use_atoms_filter(use_2)
if ok:
new = utils.form_new_clauses(
use_head_body_map,
incremental_kernel_search=True)
#new_addition.extend(new)
old_use_dict, improvedit = gl.use_dict, False
for c_ in tried_kernel_clauses:
partial_rules_copy = [
rule for rule in partial_rules if rule != c
]
partial_rules_copy.extend(new)
partial_rules_copy.append(c_)
utils.analyze_use_try(partial_rules_copy, [],
incremental_search=True)
(model_, pos, negs, optimization_score) = asp.ind(
kernel_generalization=True,
incremental_search=True)
#(lastmodel,optimization_score) = asp.ind(kernel_generalization=True,incremental_search=True)
#(pos,negs,_,_,_) = get_score(lastmodel,target_score)
#(all_solutions_,pos,negs) = asp.ind(find_all_optimal=True,
# incremental_search=True,
# opt=optimization_score)
#if pos >= pos_covered and negs <= negs_covered:
new_score = pos - negs - average_size_(model_)
if new_score > best_score:
#all_solutions = all_solutions_
model = model_
best_score = new_score
improvedit = True
print('improved it!')
#break
if not improvedit:
gl.use_dict = old_use_dict
"""
if not improvedit:
gl.use_dict = old_use_dict
if pos < pos_covered:
pos_covered = pos
if negs > negs_covered:
negs_covered = negs
"""
#if optimization_score != current_opt_score:
if True:
new_partial_rules, added = [], []
#for solution in all_solutions:
for solution in [model]:
if len(solution) > 1:
(use_2, use_3) = functs.split_use_2_3(solution)
(ok, use_head_body_map
) = functs.head_body_use_atoms_filter(use_2)
if ok:
new = utils.form_new_clauses(
use_head_body_map,
incremental_kernel_search=True)
for c in new:
if not c.as_string in added:
new_partial_rules.append(c)
added.append(c.as_string)
else:
msg = 'Found a solution use(i,j) atom with no corresponding use(i,0) atom'
raise excps.Use_2_HeadNotAbducedException(
msg, gl.logger)
partial_rules = new_partial_rules
tried_kernel_clauses.extend(partial_rules)
stop = 'stop'
kernel_loop_counter += 1
if False:
found_solution = True
else:
pass
def incremental_search(**kwargs):
import utils,core,asp,functs,excps
gl = core.global_vals
gl.use_dict = {}
is_new_example = kwargs['is_new_example']
retained,new,refined = kwargs['retcl'],kwargs['newcl'],kwargs['refcl']
prior_theory = [x for y in [retained,new,refined] for x in y]
target_score = gl.current_example_object.positive_count
if is_new_example:
utils.generate_kernel(**kwargs)
# revise the clauses in the prior theory one by one first
for c in prior_theory: # TODO
pass
var_kernel = gl.current_var_kernel
# =====================================
generate_all_seed_rules(var_kernel,gl)
# =====================================
found_solution = False
#current_opt_score = float('Inf')
#pos_covered = 0
#negs_covered = float('Inf')
best_score = 0
tried_seed_rules = []
while not found_solution:
partial_rules,tried_kernel_clauses = [],[]
kernel_loop_counter = 0
for c in var_kernel:
if 'terminatedAt' in c.as_string:
stop = 'stop'
tried_kernel_clauses.append(c)
partial_rules.append(c)
utils.analyze_use_try(partial_rules,[],incremental_search=True)
if kernel_loop_counter == 0:
(model,seed_rule,pos,negs,optimization_score) = get_seed_rule(tried_seed_rules,
var_kernel,gl)
tried_seed_rules.append(seed_rule)
print('New seed rules: ')
for r in tried_seed_rules:
print(r.as_string)
else:
(model,pos,negs,optimization_score) = asp.ind(kernel_generalization=True,
incremental_search=True)
partial_rules.extend(tried_seed_rules)
#(all_solutions,pos,negs) = asp.ind(find_all_optimal=True,
# incremental_search=True,
# opt=optimization_score)
#new_score = pos - negs - average_size(all_solutions)
new_score = pos - negs - average_size_(model)
if new_score > best_score:
best_score = new_score
print('pos,negs,average_size,best_score: ',pos,negs,average_size_(model),best_score)
#if pos >= pos_covered and negs <= negs_covered:
if False:
pos_covered,negs_covered = pos,negs
else:
# At this point, including any one solution from all_solutions reduces
# the current score. Hence, we will use one such solution as a new
# addition and search each tried kernel clause again, in order to retrieve
# the previous score (or even improve it)
#new_addition,added = [],[]
"""
for solution in all_solutions:
if len(solution) > 1:
(use_2,_) = functs.split_use_2_3(solution)
(ok,use_head_body_map) = functs.head_body_use_atoms_filter(use_2)
if ok :
new = utils.form_new_clauses(use_head_body_map,incremental_kernel_search=True)
for n in new:
if not n.as_string in added:
new_addition.append(n)
added.append(n.as_string)
"""
(use_2,_) = functs.split_use_2_3(model)
new = []
if len(use_2) > 1:
(ok,use_head_body_map) = functs.head_body_use_atoms_filter(use_2)
if ok :
new = utils.form_new_clauses(use_head_body_map,incremental_kernel_search=True)
#new_addition.extend(new)
old_use_dict,improvedit = gl.use_dict,False
for c_ in tried_kernel_clauses:
partial_rules_copy = [rule for rule in partial_rules if rule != c]
partial_rules_copy.extend(new)
partial_rules_copy.append(c_)
utils.analyze_use_try(partial_rules_copy,[],incremental_search=True)
(model_,pos,negs,optimization_score) = asp.ind(kernel_generalization=True,
incremental_search=True)
#(lastmodel,optimization_score) = asp.ind(kernel_generalization=True,incremental_search=True)
#(pos,negs,_,_,_) = get_score(lastmodel,target_score)
#(all_solutions_,pos,negs) = asp.ind(find_all_optimal=True,
# incremental_search=True,
# opt=optimization_score)
#if pos >= pos_covered and negs <= negs_covered:
new_score = pos - negs - average_size_(model_)
if new_score > best_score:
#all_solutions = all_solutions_
model = model_
best_score = new_score
improvedit = True
print('improved it!')
#break
if not improvedit:
gl.use_dict = old_use_dict
"""
if not improvedit:
gl.use_dict = old_use_dict
if pos < pos_covered:
pos_covered = pos
if negs > negs_covered:
negs_covered = negs
"""
#if optimization_score != current_opt_score:
if True:
new_partial_rules,added = [],[]
#for solution in all_solutions:
for solution in [model]:
if len(solution) > 1:
(use_2,use_3) = functs.split_use_2_3(solution)
(ok,use_head_body_map) = functs.head_body_use_atoms_filter(use_2)
if ok :
new = utils.form_new_clauses(use_head_body_map,incremental_kernel_search=True)
for c in new:
if not c.as_string in added:
new_partial_rules.append(c)
added.append(c.as_string)
else:
msg = 'Found a solution use(i,j) atom with no corresponding use(i,0) atom'
raise excps.Use_2_HeadNotAbducedException(msg,gl.logger)
partial_rules = new_partial_rules
tried_kernel_clauses.extend(partial_rules)
stop = 'stop'
kernel_loop_counter += 1
if False:
found_solution = True
else:
pass
def incremental_kernel_search(**kwargs):
gl.use_dict = {}
is_new_example = kwargs['is_new_example']
retained,new,refined = kwargs['retcl'],kwargs['newcl'],kwargs['refcl']
prior_theory = [x for y in [retained,new,refined] for x in y]
positive_count = gl.current_example_object.positive_count
if is_new_example:
generate_kernel(**kwargs)
# revise the clauses in the prior theory one by one first
for c in prior_theory: # TODO
pass
# generate new clauses from the Kernel Set, using iterative deepening
# on the subsets of the Kernel Set, until a solution is found.
#------------------------------------------------------------------------------
# TODO: This strategy can be modified to return approximate solution. This
# may be useful in cases where large effort is required to obtain a correct
# hypothesis, or in cases where a correct hypothesis does not exist (noise).
# A straightforward strategy towards this is to keep the best hypothesis found
# within a max_iterations bound.
#-------------------------------------------------------------------------------
var_kernel = gl.current_var_kernel
found_solution = False
already_found = []
for i in range(1,len(var_kernel)+1):
i_subsets = itertools.combinations(var_kernel, i)
for subset_ in i_subsets:
subset = list(subset_)
analyze_use_try(subset,[])
out = asp.ind(kernel_generalization=True)
#==============================================================================
# Test code. Try to add what you get from each subset to the bk in an effort
# to further reduce solving time.
(use_2,use_3) = functs.split_use_2_3(out)
(ok,use_head_body_map) = functs.head_body_use_atoms_filter(use_2)
if ok :
new = form_new_clauses(use_head_body_map)
if new != []:
already_found.extend(new)
n = '\n\n'.join(map(lambda x: x.as_string_with_var_types ,already_found))
write_to_file(gl.helper, n)
else:
msg = 'Found a solution use(i,j) atom with no corresponding use(i,0) atom'
raise excps.Use_2_HeadNotAbducedException(msg,gl.logger)
#===============================================================================
pos,negs,score = get_score(out)
print(str(i)+'-subsets',pos,negs,score)
if score == positive_count:
#------------------
print('Found it')
sys.exit()
#------------------
found_solution = True
(use_2,use_3) = functs.split_use_2_3(out)
(ok,use_head_body_map) = functs.head_body_use_atoms_filter(use_2)
if ok :
new = form_new_clauses(use_head_body_map)
else:
msg = 'Found a solution use(i,j) atom with no corresponding use(i,0) atom'
raise excps.Use_2_HeadNotAbducedException(msg,gl.logger)
(retained,revisable) = functs.filter_retained_(use_3,gl.use_dict)
update_support(retained)
break # stop searchin i-level subsets
if found_solution:
break # stop searching subsets
if not found_solution:
print('unsat at kernel generalization')
sys.exit()
else: # TODO
pass
def revise(**kwargs):
hs,scs,cls = kwargs['heuristic_search'],kwargs['set_cover_search'],kwargs['clause_level_search']
special_search = scs or cls
if not special_search:
gl.use_dict = {}
is_new_example = kwargs['is_new_example']
debug_mode = kwargs['debug'] if 'debug' in kwargs else False # keep this optional
retained,new,refined = kwargs['retcl'],kwargs['newcl'],kwargs['refcl']
if debug_mode:
import debug_utils
prior_theory = debug_utils.py_load_from_file()
else:
prior_theory = [retained,new,refined]
prior_theory_ = [x for y in prior_theory for x in y]
if is_new_example:
#retained.extend(refined) # that's a patch because due to a bug previously refined clauses are left out
generate_kernel(**kwargs)
if gl.runargs["kernel-set-only"]: # We only need to generate a kernel set
sys.exit()
var_kernel = gl.current_var_kernel
if 'search_subsets' in kwargs:
search_kernel_by_subsets(prior_theory_)
elif 'incremental_solve' in kwargs and kwargs['incremental_solve']:
(solution,use_2,use_3) = incremental_solve(var_kernel,prior_theory_)
else:
analyze_use_try(gl.current_var_kernel,prior_theory_)
out = asp.ind(kernel_generalization=True)
(use_2,use_3) = functs.split_use_2_3(out)
(ok,use_head_body_map) = functs.head_body_use_atoms_filter(use_2)
if ok :
new = form_new_clauses(use_head_body_map)
else:
msg = 'Found a solution use(i,j) atom with no corresponding use(i,0) atom'
raise excps.Use_2_HeadNotAbducedException(msg,gl.logger)
(retained,revisable) = functs.filter_retained_(use_3,gl.use_dict)
update_support(retained)
else: # re-seeing past example
#prior_theory = [retained,refined]
# Only analyze the new clauses generated last.
# Update This is messy...Analyze it all, there is no serious overhead
#analyze_use_try([],new,preserve=[x for y in prior_theory for x in y])
analyze_use_try([],prior_theory_)
out = asp.ind(recheck_hist_memory=True)
(_,use_3) = functs.split_use_2_3(out) # no use/2 here
(retained,revisable) = functs.filter_retained_(use_3,gl.use_dict)
revisable_ = []
for (clause_index,clause) in revisable:
#import debug_utils
#debug_utils.check_on_previous()
revcl = functs.form_revised_(clause_index,clause,gl.use_dict,use_3)
revcl = structs.Clause(revcl,gl)
revisable_.append((clause,revcl))
incorrects = [inx[0] for inx in revisable_]
specialized = []
for (incorrect,one_solution) in revisable_ : # need to find optimal refinement
optimal_ref = one_solution
check_prior = [z for z in prior_theory_ if not z in incorrects]
if search_more(initsol=one_solution,parent=incorrect,prior=check_prior):
other_init_solutions = [x[1] for x in revisable_ if x[1] != one_solution]
check_prior.extend(other_init_solutions)
optimal_ref = get_optimal_refinement(initsol=one_solution,
parent=incorrect,
prior=check_prior)
if isinstance(optimal_ref,list): # then it went through get_optimal_refinement because that returns a list
specialized.extend(optimal_ref)
else:
if isinstance(optimal_ref,structs.Clause): # then it's the initial solution, no support update yet
update_support((incorrect,optimal_ref))# this must be passed as a (parent,child) tuple
specialized.append(optimal_ref)
#d = dict(zip(incorrects,specialized))
if is_new_example:
(n,r,s) = (new,retained,specialized)
else:
(n,r,s) = ([],retained,specialized)
updateprior(n,r,s)
return (n,r,s)
else:
if kwargs['set_cover_search']:
return set_cover_search(**kwargs)
elif kwargs['heuristic_search']:
return heuristic_search(**kwargs)
else:
#return incremental_kernel_search.incremental_search(**kwargs)
return incremental_kernel_search(**kwargs)